Limitations & Honest Critiques

Accepted as valid. Mitigated as follows:

• Every substantive claim cites a primary source — Commissioner of Lobbying Registry, Ethics Commissioner reports, Auditor General reports, Canada Gazette, Federal Court records, Parliamentary committee testimony transcripts, House of Commons Journals. No secondary reporting is cited as authority.
• The investigation makes no claims about individual guilt or intent. The cross-axis tenure-overlap scores are structural indicators, not judgments. The structural finding is explicitly: no oversight body has jurisdiction over cross-axis patterns — a claim about institutional architecture, not about any specific person.
• Rejected AI outputs are documented. See the LIRIL roadmap's 4 red REJECTED rows, including R6's triple-rejection for long-form hallucinations. The review gate catches errors; review failures are logged.
• The scope is framing + evidence aggregation, not original reporting. What this site does is connect already-published primary sources across axes that Canadian media covers individually. That connection is what a structural investigation produces.

Accepted as valid. Addressed structurally as follows:

• The tenure-overlap marks are computed, not chosen. The marked-state criteria per axis are published in the axis dossier JSON files (tenure overlap ≥ 4.0 threshold with canonical-event list). Anyone can re-run the computation and get the same marks — deterministic, from the published criteria.
• Investigations span multiple parties and multiple eras. Marked actors include Liberal cabinet members (Trudeau, LeBlanc, Lametti, Sajjan, Freeland, etc.), Conservative-era actors (Harper, Poloz), and unelected bodies (PHAC, OSFI, Bank of Canada governance). The axes themselves (arms, Phoenix, MAID, judicial, Indigenous, CSIS oversight) touch multiple governments.
• Rejection of AI suggestions is red-flagged on the roadmap. R5-Q3 (LIRIL advised VPN/anonymous-email for campaigns — rejected for contextual wrongness) and R6 triple-rejection (tracking pixels, AI chatbot, long-form hallucination) show the review gate catching the AI's own biases. The rejections are the transparency, not a failure.
• Every marked actor's page links to contrary evidence when available. Wilson-Raybould's dual-axis page explicitly states the finding "does not allege misconduct" — it identifies the structural oversight gap.

The most important critique. Addressed as follows:

• Every analytical artifact is SHA-256 with published file hashes. 15+ SHA-256 file hashes published. Any claim's data file hash can be recomputed and compared to the published receipt chain. Tampering breaks the chain visibly.
• The underlying dossiers are committed JSON. Every accountability axis has a committed decision-maker JSON file in the public site repo. You can fork the repo, audit every number, and re-run the computation yourself.
• LIRIL-authored content is attributed and verifiable. Where LIRIL drafted a letter body (R3, R4, R5 campaigns — climate ECCC, CSC PBO, procurement PSPC, central banking OSFI), the consultation record with the exact prompt and response is committed at data/liril_consultation_r{N}.json — self-sha256 anchored.
• Hallucination testing is documented. 9/9 fact-retrieval probes verified correct (Bill C-7 year, Phoenix launch year, MCC name, MAID year, Pi to 4 decimals, etc.). R6 triple-rejection documents the known failure mode for long-form generation. See for the current probe score.